A PHYSICS-INFORMED NEURAL NETWORK SURROGATE MODEL FOR TIDAL SIMULATIONS

Abstract

Notable advancements in computational power has facilitated the utilization of intricate numerical methods in flood modeling in recent years. Hydrodynamic modelling approaches to predict flood inundation are robust compared to empirical approaches, which are solely based on the statistical patterns of hydrological variables obtained from observed data. Despite the benefits of numerical flood modelling tools, significant computational costs of implementing such models at high spatio-temporal resolutions have limited their applications. Data-driven machine learning (ML) models, designed to learn the underlying governing equations of the numerical models provide a computationally robust and fast alternative to the existing flood prediction models. However, these ML-based models often struggle to generalise for ‘small-data’ regime tasks, which are common to simulation-based tasks in fluid dynamics. In this study, to overcome extrapolation and over-fitting challenges of data-driven surrogate models, a Physics-Informed Neural Network model is adopted for surrogate modelling of a hydrodynamic simulator developed to investigate wave characteristics and tidal dynamics in the English Channel. A novel approach to encoding the conservation of mass into a deep learning model is introduced by including additional terms in the optimisation criterion, acting to regularise the model, avoid over-fitting and produce more physically consistent predictions by the surrogate. The model outlined improved performance by 10-20% across a range of metrics compared to a data-driven alternative.